About Embodied Science
Building intelligence that learns like a human
Embodied Science is a research lab building general-purpose embodied AI systems that can understand the physical world, learn new skills from observation, and act safely in real environments.
We believe the next breakthrough won't come from scaling task-by-task training. It will come from foundation models for action: policies that can learn quickly, generalize broadly, and improve with experience.
Teach a system by showing it real demonstrations and real-world footage so it can acquire new skills with minimal trial-and-error.
Skills should transfer across environments, objects, and variations rather than collapse outside the training distribution.
Embodied AI needs to be reliable under uncertainty. We prioritize constraints, monitoring, and predictable behavior; especially when systems operate around people.
- Measure what matters: grounded evals, real success rates, real-world robustness
- Trustworthy by design: safety and reliability are product requirements
- Generalization first: capabilities that transfer beat-out brittle demos
Research
We develop algorithms, training methods, and evaluation frameworks for embodied foundation models. We aim for fast skill acquisition, robust generalization, and safe interaction.
Applied partnerships
We work with select teams to turn research into deployed capability: data + evaluation pipelines, model development, and end-to-end systems for embodied and multimodal AI.
Work with us
We'd love to talk if you're building beyond narrow task training in robotics, autonomy, or physically grounded AI.
Publication
arXiv:2504.07524 · April 2025
GRAID: Generalized Region-based Annotations for Image Descriptions using Qualitative Spatial Reasoning
Karim Elmaaroufi et al.
